Construction of row column factorial designs

Transcription

1 J. R. Statist. Soc. B (2019) Construction of row column factorial designs J. D. Godolphin University of Surrey, Guildford, UK [Received February Final revision October 2018] Summary. The arrangement of 2 n -factorials in row column designs to estimate main effects and two-factor interactions is investigated. Single-replicate constructions are given which enable estimation of all main effects and maximize the number of estimable two-factor interactions. Constructions and guidance are given for multireplicate designs in single arrays and in multiple arrays. Consideration is given to constructions for 2 n t fractional factorials. Keywords: Replicates Confounding; Fractional factorials; Generator; Interaction; Orthogonal designs; 1. Introduction When conducting factorial experiments it is often necessary to incorporate blocking factors in the design. In experiments involving animals, litters are commonly used as blocks. In industrial experiments, typical blocking factors include site, batches of raw material and time period. Accommodating small blocks is especially challenging and it is not always possible to set up a design without confounding some effects of interest. This work concerns the construction of 2 n factorial designs with double confounding, i.e. with two forms of blocking. Designs can be represented as one or more rectangular arrays with rows and columns corresponding to blocking factors. Such an experiment with batches of raw material and machines as blocking factors is described in chapter 9 of Hinkelmann and Kempthorne (2005). Recent applications are found in Dash et al. (2013), who described microarray experiments with dye and microarray as blocking factors, and Datta et al. (2017), who outlined doubly confounded factorial designs for experiments involving livestock. Early work on factorial designs with double confounding includes Yates (1937a, b) and Rao (1946), with constructions involving quasi-latin squares. However, despite this early interest, there has been relatively little work on factorial designs with two forms of blocking. By contrast, factorial designs with one form of blocking have been extensively investigated. In particular, a number of references have formulated confounding schemes in terms of finite Abelian groups. See for example Bailey (1977, 1985), Dean and Lewis (1980) and Kobilinsky (1985). Designs for two-factors, based on Youden squares and Græco-Latin squares, are given in Preece (1966, 1971). Other work focusing on two-factors includes Dean and Lewis (1992), who studied designs with two or more blocking factors, Williams and John (1996), who adapted the simulated annealing algorithm of John and Whitaker (1993) to obtain an algorithm that Address for correspondence: J. D. Godolphin, Department of Mathematics, University of Surrey, Guildford, GU2 7XH, UK. j.godolphin@surrey.ac.uk 2018 Royal Statistical Society /19/81000

2 2 J. D. Godolphin produces designs in multiple arrays, and Gopinath et al. (2018), who gave some designs which enable main effects to be estimated with full information. More recent work that is specific to designs with factorial treatment structure that is not limited to two factors, and with double confounding, includes John and Lewis (1983), who used a cyclic approach to arrange a factorial replicate in a row column array. Here attention is not restricted to factors at two levels. Choi and Gupta (2008) produced several designs by selecting subsets of factorial effects to be confounded with blocks. For n 9, Dash et al. (2013) gave designs for s replicates of a 2 n -factorial in a 2 2 sn 1 array that provide estimates of all main effects and all two-factor interactions. Cheng and Tsai (2013) built on work of Patterson and Bailey (1978) to give the structure for a design key for a 2 n -factorial arranged in a rectangular array. Features of the design key that are necessary for estimation of main effects are highlighted. Wang (2017) gave constructions for a single replicate of a 2 n -factorial in a rectangular array which enable estimation of all main effects. Guidance is given for construction of multiplearray designs to obtain all two-factor interactions. These require a relatively large number of replicates. In this work, constructions are developed for designs comprising 2 n -factorials arranged in 2 p 2 q rectangular arrays with p+q n, from which all main effects and the maximum number of two-factor interactions are estimable. It is assumed that both forms of blocking are incomplete, i.e. n<max{p, q}. A systematic construction approach uses generating matrices, which essentially contain generating sets for two Abelian groups. These generating matrices are design keys for producing rectangular arrays containing 2 p+q n replicates of a 2 n -factorial. The approach that is used places the emphasis on which main effects and two-factor interactions are estimable, rather than on the selection of specific interactions to be confounded with rows and columns. The work expands on that of Cheng and Tsai (2013), providing a body of templates for generating matrices to give designs covering.p, q, n/ combinations that are likely to be of practical use. The paper is structured with notation and preliminary work on generating matrices in Section 2. In Section 3 constructions are developed for single-replicate arrays which provide estimates of all main effects and of the largest possible number of two-factor interactions. Section 4 covers designs comprising multiple arrays each containing a single replicate. In those cases where not all main effects and two-factor interactions can be estimated from one array, two or more arrays are used to produce a design in which each array contains some confounded effects of interest, with every required effect being estimable from at least one array. Arrays containing more than one replicate are investigated in Section 5. A brief investigation into constructions involving fractional factorials is contained in Section 6. There is some flexibility in many of the the constructions, enabling practitioners to tailor a design to give greater emphasis to specific effects. Also, guidance is given on adaptation of generating matrices to accommodate the situation in which only a subset of the two-factor interactions are of interest. Comparisons are made between designs that are obtained by the generator matrices developed and those available elsewhere in the literature. 2. Preliminaries Basic notation and concepts are consistent with chapters 7 9 and 13 of Hinkelmann and Kempthorne (2005). The factors of a 2 n -experiment are denoted A 1, A 2, :::, A n. Factor A i has levels x i {0, 1}.Ifx i = 0, then A i is described as being at low level and, if x i = 1, A i is described as being at high level. In examples where n is given, the factors are labelled A, B, :::for simplicity, rather than A 1, A 2, ::::

3 Row Column Factorial Designs 3 A treatment combination is represented as a x 1 1 ax 2 2 :::ax n n, where a0 i = 1 and a1 i = a i, and equivalently as the vector.x 1, x 2, :::, x n /. By convention, a x 1 1 ax 2 2 :::ax n n also denotes the true response. The treatment combination with all factors low is given by.1/ and.0, 0, :::,0/. A replicate involves a run in each of the 2 n treatment combinations. The vector representations of the treatment combinations are exactly the 2 n vectors in n-dimensional Euclidean space over GF(2): the Galois field of order 2. Two treatment combinations.x 1,1, :::, x 1,n / and.x 2,1, :::, x 2,n /,say,can be added to give.x 1,1 + x 2,1, :::, x 1,n + x 2,n /. A factorial effect is a contrast which partitions the treatment combinations into two sets of size 2 n 1. Combinations in the first set have an odd number of factors of the effect at low level and combinations in the second set have an even number at low level. The factorial effect is the subtraction of the average of the true responses of the first set from the average of the true responses of the second. For example, for a 2 3 -factorial, with factors A, B and C, the linear combinations corresponding to the main effect of A, also denoted A, and the interaction between A and B, denoted AB,are A = 4 1 {.1/ + a b + ab c + ac bc + abc}, AB = 4 1 {.1/ a b + ab + c ac bc + abc}: The 2 n 1 factorial effects comprise a set of mutually orthogonal contrasts in the 2 n true responses. This work concerns estimation of the n main effects and n.n 1/=2 two-factor interactions. Unless otherwise stated, interactions of three or more factors will be assumed negligible and, for brevity, interaction will be taken to mean a two-factor interaction Generating sets Consider a set of treatment combinations, T ={τ 1, :::, τ t }, which are not necessarily all distinct, where τ i =.x i,1, :::, x i,n /. Adding members of all 2 t subsets of T generates 2 t treatment combinations, which again are not necessarily distinct, each of form.σ t i=1 φ ix i,1, :::, Σ t i=1 φ ix i,n /, where φ i {0, 1} and calculations are conducted in GF(2). The set T can be summarized by a t n generating matrix: x 1,1 x 1,2 ::: x 1,n x 2,1 x 2,2 ::: x 2,n G = : : : : x t,1 x t,2 ::: x t,n The distinct treatment combinations that are generated form an Abelian group with cardinality 2 ρ, where ρ = rank.g/. In particular, for t n, the 2 t treatment combinations comprise 2 t n complete replicates of the 2 n -factorial if and only if ρ = n. To illustrate these properties, we consider two sets of four treatment combinations from a 2 3 -factorial: (a) set I, τ 1 =.1, 1, 0/, τ 2 =.1, 1, 0/, τ 3 =.1, 1, 1/ and τ 4 =.0, 1, 1/; (b) set II, τ 1 =.1, 1, 0/, τ 2 =.1, 1, 0/, τ 3 =.1, 1, 1/ and τ 4 =.0, 0, 1/. We denote the generating matrices for the sets by G I and G II respectively: G I =,

4 4 J. D. Godolphin G II = ; rank.g I /=3 and rank.g II /=2. Thus, set I generates two complete replicates of the 2 3 -factorial. The 16 treatment combinations that are generated by set II constitute four copies of the principal half-replicate with defining contrast AB, namely four copies of each of (1), ab, abc and c Row column designs Treatment combinations of a 2 n -factorial are allocated to the 2 p+q experimental units of a 2 p 2 q array, where p and q are positive integers with p + q n, to give a row column design. Without loss of generality, p q. It will be assumed throughout that q<n, so both forms of blocking are necessarily incomplete. The single-array model is y ijk = μ + τ i + ρ j + γ k + ɛ ijk :.2:1/ Here, y ijk is the observation resulting from application of the ith treatment combination to the experimental unit in the jth row and kth column. The overall mean is μ and τ i, ρ j and γ k are the effects of the ith treatment combination, the jth row and the kth column. The error terms ɛ ijk are assumed to be uncorrelated, all with variance σ 2. In the constructions developed, treatment combinations are allocated to experimental units by means of generating sets. The first column and first row of an array will be termed the principal column and principal row. These contain the treatment combinations that are generated by sets of p and q treatment combinations with generating matrices G c and G r. Treatment combination (1) is allocated to the experimental unit in the intersection of the principal column and row, i.e. in the first column of the first row. The remaining treatment combinations from G c and G r are allocated to the principal column and principal row in any order. For 2 j 2 p and 2 k 2 q, the treatment combination in row j and column k of the array is obtained by adding the vectors for treatment combinations in row j and column k of the principal column and row. In every case, once the array has been constructed, rows and columns are randomized before the design is used. With this construction process, the 2 p+q entries in the array are exactly those generated by the set of p + q treatment combinations in G c and G r. The array can be represented by a partitioned matrix G of order.p + q/ n, which is termed an array generator matrix (AGM): 2.3. Estimability conditions on G From discussion in Sections 2.1 and 2.2, the 2 p 2 q array that is obtained from G contains 2 p+q n replicates of the 2 n -factorial if and only if rank.g/ = n. It is desirable for a design to be binary with regard to both blocking factors. This occurs if and only if the 2 p treatment

5 Row Column Factorial Designs 5 combinations of the principal column are distinct and the 2 q treatment combinations of the principal row are distinct. The condition on G to achieve these properties is as follows. Condition 1. The AGM has rank.g/ = n, rank.g c / = p and rank.g r / = q. Unless stated otherwise, all matrix manipulations are conducted in GF(2). The rank of an AGM, or submatrix, is determined by reduction of the matrix to row echelon form, working modulo 2. We now examine properties of three designs, each for a replicate of a 2 5 -factorial in a 4 8 array, generated by using different AGMs. The purpose is to highlight AGM features which relate to estimability of main effects and interactions. These features will inform a construction approach for arrays with various (p, q, n) combinations Example 1 Designs D1, D2 and D3 are constructed from AGMs G 1, G 2 and G 3 : The AGMs are all full rank. Therefore, each satisfies condition 1 and yields a complete replicate of the 2 5 -factorial in a array. From the first two rows of G 1, the principal column of D1 is generated by abcd and abe and so contains treatment combinations (1), abcd, abe and cde. Likewise, from the last three rows of G 1, the principal row contains (1), acd, bce, abde, be, abcde, c and ad, giving the array D1 =.1/ acd bce abde be abcde c ad abcd b ade ce acde e abd bc : abe bcde ac d a cd abce bde cde ae bd abc bcd ab de ace Designs D2 and D3 are obtained from G 2 and G 3 in the same way:.1/ ace bcd abde bce ab de acd abcd bde a ce ade cd abce b D2 = ; abe bc acde d ac e abd bcde cde ad be abc bd abcde c ae D3 =.1/ ac abd bcd bce abe acde de ade cde be abce abcd bd c a : bde abcde ae ce cd ad abc b ab bc d acd ace e bcde abde Inspecting the design array for D1, each factor is high in exactly half the treatment combinations in each row and column. Thus all main effects are estimable. Next consider the interaction AB. Half the treatment combinations in each row have an even number of A and B high. However, each column either has all treatment combinations with an even number of A and B high or all with an odd number of A and B high. The same property is noted for CD. Thus AB and CD are confounded with columns. In a similar way, AD and BE are confounded with rows. The other interactions have half the treatment combinations in each row and each column with

6 6 J. D. Godolphin an even number of the two factors high. Thus AC, AE, BC, BD, CE and DE are estimable but AB, AD, BE and CD are not. These properties can be deduced directly from G 1 =.G1c T GT 1r /T, where M T denotes the transpose of matrix M. Each column of G 1c and of G 1r contains at least one element of 1, with consequence that each treatment is high in exactly half the experimental units in the principal column and in the principal row and therefore in each column and row of the array. Thus, in the linear combination of observations for a main effect, the row and column effects add out and the effect is estimable. Considering AB, the first and second columns of G 1c are the same, resulting in all treatment combinations in the principal column having A and B both high or both low. All have positive sign in the AB-effect. As the remaining columns are cosets of the principal column, each column has all treatment combinations with an even number of A and B high or all with an odd number of A and B high. Thus, all treatment combinations in a column have the same sign in the AB-effect and AB is completely confounded with columns. The third and fourth columns of G 1c are also the same and CD is not estimable. Likewise, AD and BE are inestimable because of repeated columns in G 1r. Conversely, consider AC. The corresponding columns of G 1c are different as are those of G 1r. Thus, in both the principal column and the principal row, exactly half the treatment combinations have positive sign in the AC-contrast. This property extends through all columns and rows, leading to estimability of the effect. In design D2, all main effects are estimable since each column of G 2c and G 2r contains at least one element of 1. Submatrix G 2c is the same as G 1c and, by the same reasoning as above, AB and CD are not estimable. The columns of G 2r are all different and so all other interactions are estimable. The notable feature of G 3 is that the third column of G 3c contains only 0s, with the consequence that C is low in all treatment combinations in the principal column of design D3. Thus, C is confounded with columns and is not estimable. All other main effects are estimable as are all interactions except DE. Properties that are highlighted in example 1 prompt general conditions on AGMs relating to estimability of main effects and interactions. As seen in example 1, an effect is either estimable with full information or is completely confounded with blocks and is inestimable. Some notation is introduced to facilitate statement of the conditions. Let X U be the set of non-zero vectors of length U over GF(2). Then X U =2 U 1. For example, ) ) ) ) ) ) )} X 3 = {( 1 1 1, ( 1 1 0, ( 1 0 1, ( 0 1 1, ( Condition 1 can be supplemented by a second condition., ( 0 1 0, ( Condition 2(a). Each column of G c or G r is a vector of X p or X q respectively. Together, conditions 1 and 2(a) are necessary and sufficient for G to be the AGM for a 2 p 2 q array containing 2 p+q n replicates of a 2 n -factorial, from which all main effects are estimable. Condition 2(a) was given in Cheng and Tsai (2013). Designs D1 and D2 of example 1 have AGMs which satisfy conditions 1 and 2(a). These designs have different properties with respect to estimation of interactions and we seek an alternative to condition 2(a) which, with condition 1, relates to estimation of all main effects and to maximizing the number of estimable two-factor interactions. From discussion in example 1, for an AGM satisfying condition 1, an interaction is estimable if and only if the corresponding columns in G c and G r are distinct. Thus, a pair of common columns in G c corresponds to an interaction that is not estimable. The set X p has cardinality 2 p 1. The number of columns of :

7 Row Column Factorial Designs 7 G c can be expressed uniquely as n = α.2 p 1/ + β where α = [n=.2 p 1/], with [ ] denoting the integer part of, and β {0, 1, :::,2 p 2}. By manipulation of binomial coefficients, an upper bound for the number of pairs of distinct vectors in a set of n vectors from X p is ( ) ( ) n ω = αβ.2 p α 1/ :.2:2/ 2 2 This is also an upper bound for the number of estimable interactions from an array and reduces to. n 2/ when n 2 p 1. The bound (2.2) prompts an alternative to condition 2(a), labelled condition 2(b), which concerns estimability of both main effects and interactions. Condition 2(b). G c is formed from α + 1 copies of each of β vectors of X p and of α copies of each of the remaining 2 p β 1 vectors, and the columns of G r are distinct. Together, conditions 1 and 2(b) are necessary and sufficient for an AGM to yield an array from which all main effects and ω interactions are estimable. Both conditions are satisfied by design D2 of example 1 but not by D1. No arrangement of a 2 5 -factorial in a array can provide estimates of all main effects and of more than ω = 8 interactions. In general, for a single-replicate array with q>2 p p 1, then n = p + q>2 p 1 and G c will contain repeated columns. For such parameter combinations it is not possible to estimate all main effects and interactions from one array. Note that, for q 3, we have n 2q<2 q 1, indicating that X q >n, i.e. the number of distinct vectors that are available for use in G r exceeds n. 3. Single-replicate arrays We now investigate the arrangement of a single replicate of a 2 n -factorial in a 2 p 2 q array, with p + q = n. The task is to give AGMs for different (p, q) combinations which generate arrays yielding estimates of all main effects and of ω interactions. The arrays can be used as singlereplicate designs, after randomization of rows and columns, or combined to form multireplicate designs, which are covered in Sections 4 and Estimability of main effects We first focus on main effects and give constructions for AGMs which satisfy conditions 1 and 2(a) and thus yield single-replicate arrays from which all main effects are estimable. There are two cases. Theorem 1. For p+q 4, an AGM of the following form yields a replicate of a 2 p+q factorial ina2 p 2 q array enabling estimation of all main effects. (a) Case 1, pq even: (3.1a) (b) Case 2, pq odd or q even: (3.1b) where I v and J v,w denote the v v identity matrix and the v w matrix with each term 1.

8 8 J. D. Godolphin Proof. Consider the.p + q/.p + q/ matrices and where 0 v,w denotes the v w matrix, with all entries 0. All entries are in GF(2). Matrix M 1 is full rank. For q even, M 2 has row-reduced echelon form equal to M 1 and is therefore also full rank. (a) Case 1: on adding the sum of rows 1, :::, p to row i, fori = p + 1, :::, p + q, working modulo 2, the matrix G is transformed to M 1 if p is even, and to M 2 if p is odd and q even. Thus, G has full rank. (b) Case 2: likewise, on adding the sum of rows 1, :::, p to row i, fori = p + 1, :::, p + q, the matrix G is transformed to M 1 if p is odd, and to M 2 if p and q are both even. Thus, G has full rank. In both cases G has full rank, satisfying condition 1. Further, in both cases it is evident that condition 2(a) is satisfied, and the result follows. Case 1 covers all p, q pairs except those with p and q both odd, and case 2 covers all pairs except those with p even and q odd. No constructions are given for p + q 3 since no degrees of freedom would be available for estimation of σ 2. The constructions of theorem 1 have similarities to those given by Wang (2017). Case 2 always applies for p=1. The AGM for p=1 has relevance later and is covered now as a corollary. Corollary 1. For p=1 and q 3, an AGM which arranges a single replicate of a 2 q+1 -factorial ina2 2 q array from all main effects are estimable is (3.2) Example 2 Examples of AGMs for a single replicate of a 2 5 -factorial from which all main effects are estimable with (a) p = 2 and q = 3, and (b) p = 1 and q = 4 are respectively and

9 Row Column Factorial Designs 9 In many cases the arrays of theorem 1 can be improved on if interactions are also of interest. To demonstrate this, denote the design corresponding to the array of example 2, combination (a), by D4. In common with D1 and D2 of example 1, design D4 has p=2 and q =3 and satisfies conditions 1 and 2(a). Of the three designs, D2 is the only one which satisfies condition 2(b) and so achieves the bound (2.2) with eight estimable interactions. Designs D1 and D4 provide estimates of only six interactions. We now develop AGM constructions which satisfy conditions 1 and 2(b) and therefore yield arrays enabling estimation of all main effects and of ω interactions. AGMs are given for all p and q with p + q = n, where p + q 4, except for the case p = q = 2 which is considered separately Constructions for 3 p q Lemma 1. For p 3, let K p =.k i,j / be the p p matrix with k i,j = 1ifi j and k i,j = 0if i>j, and let L p =.l i,j / = K p + K p J p,p + I p. Then the columns of the p 2p matrix.k p L p / are 2p distinct columns from X p. Proof. The jth column of K p has first j elements 1 and last p j elements 0. Thus the columns of K p are all distinct and all in X p. Now consider L p. By construction, we have l i,j = p + 1 i if i j and l i,j = p + 2 i if i<j. Therefore the columns of L p are distinct and, since l p,1 = l p,2 =:::= l p,p = 1, each is a column of X p.forpodd, l 1,j = 0, for j 2, and l 2,1 = 0. For p even, l 2,j = 0, for j 3, and l 3,1 = l 3,2 = l 3,3 = 0. Therefore, each column of L p contains at least one element 0 which comes above an element of 1 and so none is a column of K p. Thus, the columns of.k p L p / are distinct columns of X p as required. The condition p 3 is needed in lemma 1 since the columns of.k p L p / are not distinct for p 2. Theorem 2. For 3 p q, an AGM of the following form arranges a replicate of a 2 p+q - factorial in a 2 p 2 q array, from which all main effects and ω interactions are estimable: where XÅ is a p.q p/ matrix such that each column of X p appears in the p n matrix.i p I p + J p,p XÅ/, α or α + 1 times, where α = [.q + 2/=.2 p 1/]. Proof. Reducing G to row echelon form, working modulo 2, gives which confirms that G is full rank, satisfying condition 1. By the construction of XÅ, every column of G c is in X p and the columns have multiplicities that are consistent with condition 2(b).

10 10 J. D. Godolphin By lemma 1, the first 2p columns of G r are distinct columns of X q. Further, the final q p columns are also distinct columns of X q and are different from any of the first 2p columns. Thus condition 2(b) is satisfied which establishes that all main effects and ω interactions are estimable from the array that is generated by G Example 3 Examples of AGMs for single-replicate arrays from which all main effects and ω interactions are estimable are given for (a) p = 3 and q = 3, and (b) p = 3 and q = 6 respectively by and All effects of interest are estimable from combination (a). By comparison, for p = q = 3, the construction of Wang (2017) gives an array with six inestimable interactions and requires three single-replicate arrays to achieve estimability of all main effects and interactions. From combination (b) all effects of interest bar AH and BI are estimable. Because of the reliance of theorem 2 on lemma 1, the construction does not apply to arrays with p<3. Separate constructions are required for p = 2 and p = Constructions for p D 2 and q 3 No array comprising a single replicate of the 2 q+2 -factorial will give estimates of all main effects and interactions because there are q columns in G, but only three vectors in X 2. The following result gives constructions which enable estimation of all main effects and maximize the number of estimable interactions. Theorem 3. For q 3, an AGM of the following form yields a replicate of a 2 q+2 -factorial in a2 2 2 q array from which all main effects and ω interactions are estimable: (3.3)

11 where E and F are the matrices E = F = ), ( ( ), Row Column Factorial Designs 11 and XÅ isa2.q 3/ matrix which is selected so that each vector of X 2 appears in the 2 n matrix.i 2 FXÅ/, α or α + 1 times, where α = [.q + 2/=3]. Proof. Reducing G to row echelon form gives which confirms that G is full rank, satisfying condition 1. By the selection of columns of XÅ, every column of G c is a vector of X 2 and the multiplicities are consistent with condition 2(b). Now consider the 3 5 top left-hand submatrix of G r : This submatrix comprises distinct vectors of X 3. Thus, the columns of G r are distinct vectors of X q. It follows that condition 2(b) is satisfied, which establishes the result Example 4 An AGM for a replicate of the 2 6 -factorial arranged in a array from which all main effects and ω interactions are estimable is (3.4) Of the 15 interactions, all are estimable except AE, BD and CF. Arrays from theorem 3 improve on arrays and constructions elsewhere in the literature, with regard to estimation of interactions. For q = 3, the array of theorem 3 has two confounded interactions. With the same parameters, the construction of Wang (2017) and arrays in Choi and Gupta (2008) each have three or more confounded interactions. However, it is noted that Choi and Gupta (2008) did not restrict attention to interactions of only two-factors Construction for p D 2 and q D 2 Theorem 3 does not apply to p = q = 2. In this case condition 2(b) cannot be satisfied since the number of columns of G r exceeds the cardinality of X 2. Therefore there is no arrangement of a

12 12 J. D. Godolphin replicate of the 2 4 -factorial in a 4 4 array which enables estimation of all main effects and of ω =5 interactions. The two arrays from theorem 1 enable estimation of all main effects and of four of the six interactions and cannot be improved on. The AGM that is obtained from theorem 1 is (3.5) In practice, a design consisting solely of this array would not be recommended for experimentation since only 1 degree of freedom is available for estimating σ 2. A two-replicate design arranged in two 4 4 arrays is considered in Section 4. Wang (2017) and Choi and Gupta (2008) also gave single-replicate designs with p = q = 2. Both designs enable estimation of all main effects. The design of Wang (2017) is equivalent to AGM (3.5) and gives estimates of four interactions, but only three interactions are estimable from the design of Choi and Gupta (2008) Constructions for p D 1 and q 3 Corollary 1 yields a single-replicate 2 2 q array from which all main effects are estimable. Technically, the AGM of corollary 1 satisfies condition 2(b) but, since the upper bound for the number of estimable interactions given by expression (2.2) is ω = 0, no interactions are estimable from the array. This is readily seen from expression (3.2), since every column of G c is identical. Single-replicate arrays can be constructed, enabling estimation of some interactions but these arrays do not enable estimation of all main effects. This is investigated in Section 4. Theorems 2 and 3 and corollary 1 give single-replicate arrays for all parameter combinations with 1 p q where q 3. For parameter combinations with 3 p q and p + q 2 p 1, the arrays of theorem 2 enable estimation of all main effects and interactions. Arrays for all the other parameter combinations enable estimation of all main effects and of ω interactions. In general, the arrays of theorems 2 and 3 have advantages over constructions of Wang (2017) which have {p.p 1/ + q.q 1/}=2 inestimable interactions. For cases where not all interactions are estimable or where, as for the design that is generated by AGM (3.5), the number of degrees of freedom that are available for estimation of σ 2 is very small, multireplicate designs are advised. These are investigated in Sections 4 and Bespoke single-replicate arrays There are situations with p 2 and q 3 in which an array providing estimates of fewer than ω interactions may be preferable to an array from theorem 2 or theorem 3. Retaining the AGM structures of theorems 2 and 3 and the requirement that columns of XÅ are vectors of X p,but relaxing the requirement regarding selection of XÅ to achieve as near equal as possible multiplicities of columns in G c can reduce ω but may be beneficial for a given experimental situation Example 5 To investigate an industrial process, six factors are identified for experimentation. A design comprising a replicate of a 2 6 -factorial arranged in a array is required. Because of the nature of the factors and process it can be assumed that there are no interactions between a subset of three factors and no interaction between a further two factors. The design that is given by theorem 3 has XÅ =.1, 1/ T and yields an array in which all main effects and 12 of the 15 interactions are estimable, with AE, BD and CF being inestimable.

13 Row Column Factorial Designs 13 In allocating factors to labels, at least one interaction of interest will be inestimable. If XÅ =.1, 0/ T is used in expression (3.3), then only 11 interactions are estimable: the inestimable interactions are AE, AF, EF and BD. Allocating the three factors that do not interact with A, E and F and the further two-factors that do not interact with B and D gives an array from which all effects of interest are obtained. 4. Multiple single-replicate arrays We now consider designs comprising a number of arrays each containing a single replicate. For an m-replicate 2 n factorial design, each replicate is arranged in a separate 2 p 2 q array where p + q = n. The model given by expression (2.1) is adjusted to incorporate replicates as an additional blocking factor: y ijkl = μ + τ i + ρ jl + γ kl + α l + ɛ ijkl, where y ijkl is the observation on application of the ith treatment combination to the experimental unit in the jth row and kth column of the lth replicate and α l is the effect of the lth replicate. Other terms are obvious extensions of terms in expression (2.1), with γ kl and ρ kl being the effects of the jth row and the kth column in the lth replicate Multiple-array designs for p 2 and q 3 For (p, q) combinations with both p 2 and q 3, individual replicates are constructed by using results from Section 3. The AGM for the ith replicate is denoted by G i, with submatrices G i c and G i r.ifp 3 and q 2p p 1, the construction of theorem 2 gives replicates which each provide estimates of all effects of interest. Otherwise, i.e. if p 3 and q>2 p p 1orifp = 2 and q 3, although every replicate yields estimates of all main effects, no single replicate enables estimation of all interactions. The columns of generator matrices from theorems 2 and 3 are rearranged to give AGMs for a sufficient number of individual replicates, so the resulting arrays together provide estimates of all interactions. Main effects are estimated with full efficiency and some interactions are partially confounded. This is now demonstrated Example 6 A two-replicate design is required for a 2 6 -factorial with each replicate arranged in a array. The AGM (3.4) is used as G 1. This provides estimates of all main effects and all interactions except AE, BD and CF. The columns of G 1 are rearranged to give G 2 : (4.1) No vertical construction lines are included in equation (4.1), since G 2 is a rearrangement of an existing AGM. The second replicate provides estimates of all effects of interest except AF, BE and CD. Thus, the design comprising both arrays enables estimation of all main effects and nine interactions with full information. The remaining six interactions are partially confounded and are estimated with 2 1 relative information.

14 14 J. D. Godolphin For an array of theorem 2 or theorem 3, any non-estimability of interactions occurs as a consequence of repetition of columns in G c. Stacking the submatrices Gc 1 and G2 c of example 6 demonstrates the property that results in full estimability: Although both submatrices necessarily have repeated columns, the 4 6 stacked matrix has no repeated columns, indicating that each interaction is estimable in at least one replicate. In general, for an m-replicate design we define S c and S r to be the mp n and mq n matrices formed by stacking the AGM submatrices: Let the AGMs for m replicates be obtained from theorem 2 or 3, possibly with rearranged columns. Then, columns of Gr i comprise n distinct vectors from X q and columns of Gc i are vectors from X p. Every main effect is estimable from every replicate. An interaction is estimable from at least one replicate if and only if the corresponding columns of S c are different. Further, denoting the ith column of Gc k by gk ci, the interaction A ia j is estimable from m i,j = m m [ p.g k ci gcj k /T.gci k gk cj / ].4:2/ k=1 p replicates, where [ ] denotes the integer part of. Calculations of equation (4.2) are evaluated in R and R p and not in GF(2). The interaction A i A j is estimated with m i,j =m relative information. The set X p has cardinality 2 p 1 and therefore there are.2 p 1/ m different possibilities for columns of S c, each obtained by stacking m, not necessarily distinct, vectors from X p. The bound (2.2) can be extended to give an upper bound for the number of interactions that are estimable from a set of m arrays. The number of columns of S c can be expressed as n = α m.2 p 1/ m + β m where α m = [n=.2 p 1/ m ] and β m {0, 1, :::,.2 p 1/ m 1}. Then, an upper bound for the number of pairs of distinct columns from the columns of S c, and hence for the number of interactions that are estimable from at least one array, is ( ) ( ) n ω m = α 2 m β m.2 p 1/ m αm :.4:3/ 2 Note that equation (4.3) reduces to equation (2.2) for m = 1. The bound leads to the following proposition.

15 Row Column Factorial Designs 15 Proposition 1. If p 2 and q 3, a design comprising log 2 p 1.n/ single-replicate arrays exists such that all main effects are estimable from every array and each interaction is estimable from at least one array. By proposition 1, for p 3, designs can be constructed in two replicates which enable estimation of all effects of interest, for n 49. Likewise, for p=2 and q 3, two and three replicates are sufficient for n 9 and n 27 respectively. In practice, it is straightforward to construct designs in log 2 p 1.n/ arrays with estimation properties that are consistent with proposition 1. For small n, it may be desirable to use a design with more than log 2 p 1.n/ single-replicate arrays, to increase the degrees of freedom that are available for estimation of σ Example 7 A multireplicate design for a 2 8 -factorial is required, with each replicate arranged as a array. By proposition 1, a design can be constructed in two single-replicate arrays, such that all main effects are estimable from both arrays and every interaction is estimable from at least one array. Using theorem 3, a possible G 1 c is ( ) Gc = : Rearranging the columns of Gc 1 in the order 1, 2, 4, 3, 8, 5, 6, 7 gives a G2 c, with the submatrices stacking as All columns of S c are distinct. The second replicate is constructed with G 2 formed by arranging the columns of G 1 in the order that is given above. All effects of interest except AE, AG, EG, BD, BH, DH and CF are estimable from the first replicate. The non-estimable interactions from the second replicate are AF, AH, FH, BC, BE, CE and DG. Thus, the two-replicate design provides estimates of all main effects and of 14 interactions with full information and of the remaining 14 interactions with 2 1 relative information. Proposition 1 and the approach of example 7 can be used to construct designs comprising multiple single-replicate arrays for all.p, q/ pairs except for p = q = 2 and p = 1. These cases are now considered Multiple-array design for p = q = 2 For p = q = 2, the submatrix G r for a single-replicate array will not have all columns distinct. Therefore, attention needs to be given to both S c and S r when constructing a multireplicate design. We use the generator matrix that is given in expression (3.5) for G 1. This first replicate provides estimates of all required effects except AB and CD. For a second replicate, switch the second and third columns of G 1 to give

16 16 J. D. Godolphin The second replicate gives estimates of all effects of interest except AC and BD. Thus, a design comprising these replicates has all main effects and two interactions estimable with full efficiency and the remaining four interactions estimable with 1 2 relative information. Note that S c and S r each contain four distinct vectors from X 4. Of the 31 degrees of freedom, eight are available for estimating σ 2. The properties of the two-replicate design can be contrasted with those of the single-replicate design with p = q = 2 in Section Multiple-array design for p D 1 and q 3 For single-replicate arrays with p = 1, the only G c submatrix that satisfies condition 2(a) is G c = 1 T n. This is reflected in the construction of corollary 1. An array that is produced by this construction gives estimates of all main effects but of no interactions. In this work, estimation of main effects is taken to have greater priority than estimation of interactions and up to this point all constructions have involved arrays from which all main effects are estimable. However, in the case of a multireplicate design with p=1, to estimate any interactions it is necessary to use some arrays which do not satisfy condition 2(a). With this adjustment, single-replicate arrays are used which allow estimation of a subset of the main effects and of the interactions. With careful selection, a design comprising a set of arrays will give estimates of all effects of interest. Lemma 2. For q 3, an AGM of the following form yields a single replicate of a 2 q+1 -factorial ina2 2 q array from which q 0 +1 main effects and.q 0 +1/.q q 0 / interactions are estimable, where q 0 {0, 1, :::, q 1}: (4.4) The proof is along the same lines as those of theorems 2 and 3 and is not included. The condition q 3 ensures that the columns of G r are distinct vectors of X q and therefore that estimable effects of interest depend entirely on G c. Main effects A 1, :::, A q0 +1 are estimable as are all interactions involving one of these factors and one of A q0 +2, :::, A n. Theorem 4. For q 3, let = log 2.q + 2/ and let the.q + 1/ matrix SÅ be formed from q + 1 distinct vectors of X. Then SÅ is the S c -matrix for a design in replicates of a 2 q+1 -factorial, each arranged as a 2 2 q array, such that each main effect and interaction are estimable from at least one replicate. Proof. From the definition of, wehave2 1 1 <q Of the vectors in X, 2 1 have the jth element 1 and 2 1 1havejth element 0, for j = 1, :::,. Therefore, in any set of q + 1 distinct vectors of X, at least one has entry 1 in the jth position for j = 1, :::,.

17 Row Column Factorial Designs 17 Thus, each row of SÅ contains at least one element of 1 and is a 1.q + 1/ vector corresponding to G c for a single-replicate array for a 2 q+1 -factorial formed from corollary 1 or lemma 2, possibly after rearrangement of columns. Let the ith row of SÅ be Gc i. The Gi r-submatrix that completes G i is obtained by rearranging the columns of the appropriate G r -submatrix of matrix (3.2) or (4.4) so that G i corresponds to matrix (3.2) or (4.4) with columns rearranged. Consider the design comprising single-replicate 2 2 q arrays with AGMs G 1, :::, G. The effects of interest that are estimable from the ith replicate are identified directly from Gc i. Since each column of SÅ contains at least one element 1, each main effect is estimable from at least one replicate. Finally, since the columns of SÅ are all different, every interaction is estimable from at least one replicate. The result follows. Note that there is no requirement for SÅ of theorem 4 to be full rank Example 8 A multireplicate design for a 2 5 -factorial is required, with each replicate arranged as a array. By theorem 4, designs can be constructed in =3 single-replicate arrays. We start by selecting five distinct vectors of X 3 to form SÅ. For example,.4:5/ Using lemma 2, and rearranging columns so that the first row of each AGM is consistent with the corresponding row of S c, gives The main effect of E is estimable from each replicate, whereas A, B and C are each obtained from two replicates and D is estimable only from the second replicate. Each interaction is estimable from at least one replicate. For example, BD is estimable from only the first replicate whereas AD is estimable from all three replicates.

18 18 J. D. Godolphin Unlike other constructions in this work, designs from theorem 4 do not have all main effects estimable from every replicate. Indeed, at most one replicate will have G c = 1 T q+1, so at least 1 replicates will not provide estimates for the full quota of q + 1 main effects. Properties of designs of theorem 4 can be identified directly from SÅ. Here, calculations are in R and R m. The ith element of 1 T S Å is the number of replicates providing an estimate of A i. Further, let the.q + 1/.q + 1/ matrix TÅ =.t ij / have t ij =.s i s j / T.s i s j /, where s i is the ith column of SÅ. Then t ij is the number of replicates giving estimates of A i A j,fori j. The numbers of main effect and interaction estimates from the replicates are 1 T S Å1 and 1 T T Å1 =2. The selection of vectors from X to form SÅ does have implications on the estimability properties of the design. This is now considered briefly and illustrated with reference to example Example 8 (continued) For the SÅ-matrix of expression (4.5), 1 T 3 S Å = /, :6/ TÅ = : The total numbers of main effect estimates and interaction estimates from the three replicates are 1 T 3 S Å1 3 =10 and 1 T 3 T Å1 3 =2=16 respectively. It is interesting to consider the effect of changing the selection of vectors of X 3 that are used in SÅ. First relabel SÅ and TÅ of expressions (4.5) and (4.6) as SÅ1 and TÅ1, and denote the design by D81. An alternative design, D82, is obtained by replacing the final column of SÅ by.1, 0, 0/ T. We denote this alternative matrix by SÅ2 with corresponding matrix TÅ2 to give ( ) 1 SÅ2 = , T 3 S Å2 = /,.4:7/ TÅ2 = : In design D82, the main effect E is estimable from only one replicate and the total number of main effect estimates is 1 T 3 S 2Å1 3 = 8 compared with the 10 estimates from D81. The number of interaction estimates from D82 is 1 T 3 T Å21 3 =2 = 18: an increase on the 16 estimates from D81. Focusing on those effects estimated with smallest relative information, from the design summaries in expressions (4.6) and (4.7) it is observed that in D81 one main effect and five interactions are estimable from only one replicate but that in D82 two main effects and four interactions have this property. Dash et al. (2013) provided designs in replicates of a 2 q+1 -factorial arranged in 2 2 q arrays for 2 n = q which enable estimation of all effects of interest. Theorem 4 gives designs with equivalent estimability properties to those of Table 3 of Dash et al. (2013), in the sense

19 Row Column Factorial Designs 19 that the numbers and distributions of independent estimates of main effects and two-factor interactions are the same Example 8 (continued) The two designs that were given in Dash et al. (2013) for three single-replicate arrays of a 2 5 -factorial both have estimability properties that are equivalent to that of a design of theorem 4, design D83 say, with SÅ3 = ( ) , T 3 S Å3 = /, TÅ3 = : Design D83 has 1 T 3 S Å31 3 = 9 independent main effect estimates and 1 T 3 T Å31 3 =2 = 18 independent interaction estimates. Two main effects and four interactions are estimable from only one replicate. In general, selection of vectors from X with a greater number of elements 1 gives greater emphasis to estimation of main effects. Selection of vectors so that the number of non-zero elements in a row of SÅ is close to.q + 1/=2 gives greater emphasis to estimation of interactions. This is demonstrated in example 8. For some values of q there are a large number of selections of q + 1 vectors from X. In such cases, a judicious choice of vectors and arrangement of these to form SÅ enables construction of bespoke designs that provide estimates of selected effects with high relative information. Conversely, for other values of q, the range of estimability properties of designs from theorem 4 is very limited. This is expressed in the following corollary to theorem 4. Corollary 2. If q = 2 k 2 for some integer k 3, then the designs of theorem 4 all have equivalent estimability properties. If q = 2 k 3, then the designs of theorem 4 can be partitioned into k types with respect to estimability properties. Proof. The estimability properties of an -array design obtained from theorem 4 are determined by the columns of SÅ.Ifq = 2 k 2or2 k 3, for k 3, then = k and SÅ is formed from distinct vectors of X k.forq = 2 k 2, matrix SÅ comprises all q + 1 = 2 k 1 vectors of X k and thus all designs have identical SÅ- and TÅ-matrices, bar a reordering. For q = 2 k 3, matrix SÅ comprises all except one of the vectors of X k. The number of 0s in the omitted column vector can be any one of 0, 1, :::, k 1, giving k estimability structures. 5. Multireplicate arrays We now consider single-array designs containing two or more replicates. Model (2.1) applies and the AGM has order.p + q/ n with n<p+ q. Further, q<n, since only arrays with incomplete blocks are considered. The design consists of 2 p+q n replicates in a 2 p 2 q array. The approach that is used is to adapt the constructions for single-replicate arrays of Section 3

20 20 J. D. Godolphin and the selection of sets of single-replicate arrays of Section 4. Fewer degrees of freedom are required for nuisance parameters, i.e. blocks or replicates, than for a design comprising several arrays with the same number of runs. Note that p 2 and that, in terms of p and q, possible values for n are p + 1, :::, p + q Multireplicate array for p 3 and q<n<pcq We give constructions satisfying conditions 1 and 2(b). Define and where K p and L p are as given in lemma 1, and let M.a/ denote the first a columns of matrix M. Theorem 5. An AGM of the form that is specified below generates a 2 p 2 q binary array containing 2 p+q n replicates of a 2 n -factorial, from which all main effects and ω interactions are estimable. (a) Case 1, n 2p: (b) Case 2, n > 2p: where XÅ is a p.n 2p/ matrix such that each vector of X p appears in the p n matrix G c α or α + 1 times, where α = [n=.2 p 1/]. The proof is by an argument similar to that of theorem 2 with the minor adjustment that it is necessary to demonstrate that G c and G r are each full rank for the design to be binary with respect to rows and columns. Since the columns of G r are distinct, estimability properties are determined by the submatrix G c.ifn 2 p 1, then the columns of G c are also distinct and the array of theorem 5 provides estimates of all effects of interest with full information. Otherwise, some interactions will be inestimable. The arrays of theorem 5 are resolvable in both rows and columns. The 2 q treatment combinations of the principal row partition the 2 n treatment combinations into 2 n q cosets. The 2 q treatment combinations in each coset occur together in 2 p+q n rows. A similar property is noted in columns. This property causes the design to be resolvable in various alternative ways with respect to rows and columns. For example, for a two-replicate design, i.e. a design with 2 p+q n = 2, the rows can be resolved into two replicates in 2 2n q 1 ways. Further, for 2 n q of these ways of resolving the rows, the individual replicates can be expressed in terms of generator matrices. Similarly, there are 2 2n p 1 ways of resolving the columns into two replicates, of which 2 n p can be given in terms of generator matrices. The property of being resolvable according to generator matrix constructions can be beneficial in the event of observation loss, as discussed in the following example.